8,300 stars on GitHub: this tool disrupts AI and could very well turn against its creators

An open source project called Heretic is gaining traction on GitHub this February 2026. Behind its technical prowess, it automatically removes the safety measures of AI models. The result: potentially uncontrollable uses, both for others and for yourself.

An tool capable of automatically removing safety measures from AI models without significantly degrading their performance

For several years, large language models have incorporated safety measures. They block illegal or dangerous requests. Until now, circumventing these protections required cumbersome and risky manual code modifications.

Heretic changes the game. The tool operates entirely automatically and only targets the alignment mechanisms. According to its creators, the model retains most of its capabilities. Tests indicate a low KL divergence, a sign of performance close to the original.

Why this technical feat poses a concrete threat in the age of autonomous AI agents

The risk goes beyond simple circumvention. In the era of agentic AI, some open source projects are already executing actions on your devices. Combined with these systems, Heretic could remove all operational limits.

In practical terms, a modified agent could automate hacking attempts or orchestrate large-scale campaigns. It would become a phishing factory capable of crafting and sending fraudulent messages on a massive scale. The damage could quickly extend beyond individual cases.

Additionally, these agents never rest. Once freed from their constraints, they exploit every available resource. In this context, the removal of barriers creates a major systemic risk for the digital ecosystem.

A danger that can also turn against the user via prompt injection attacks

The issue doesn’t only concern potential victims. By removing internal protections, you also expose your own digital environment. Prompt injection attacks are advancing and becoming more sophisticated each month.

Hackers now conceal malicious instructions within images or emails, where normally a model would block certain suspicious commands. Without these filters, your agent might execute actions against your interests through hidden instructions that are difficult to detect.

Growing popularity on GitHub, current limitations, and best practices to mitigate risks

Heretic is easy to use for those who understand the command line. The official repository already boasts more than 8,300 GitHub stars and hundreds of project forks. This visibility accelerates its spread within the technical community.

For now, the tool primarily targets small open source models. Major proprietary services do not seem directly affected. However, some powerful open models are bridging that gap, raising concerns.

In response to this evolution, you need to strengthen your security practices. Enable two-factor authentication, keep your devices updated, and deny unnecessary access. Finally, avoid any suspicious links. Daily vigilance remains your best protection in 2026.

Scroll to Top